Current ways of developing AI agents contribute to degradation of oversight

Because the abstract includes both interface barriers and skill atrophy, task accuracy alone would miss an important part of the oversight risk it describes.